Mediation Analysis

A mediator is a variable or a construct that intervene between the independent variable(s) and the dependent variable(s). This study used the approach developed by Preacher and Hayes (2008) to test for the mediation effect. Specifically, this study employed Hayes’ (2013) PROCESS macro for SPSS to compute confidence intervals for specific indirect effects based on 5,000 bootstrap samples as AMOS does not perform bootstrapping for specific but only for total indirect effects. To determine whether seeking resources and seeking challenges mediate the relationships between the antecedent and outcome variables, bootstrapping analyses were performed using methods described by Preacher and Hayes (2008) for estimating direct and indirect effects with multiple mediators.

A mediation analysis comprises three effects of X on Y: the first effect is the direct effect c’, the second effect is the indirect effect ab, and the third effect is the total effect

c (Preacher & Hayes, 2008). Hence, to test for a mediation effect or indirect effect,

according to Preacher and Hayes (2008), do not require the direct effect before testing the indirect effect. In other words, the indirect effect results from the causal influence of X on M which in turn affect Y is manifested through mean differences. With this logic, the direct effect is simply the mean difference in Y, regardless of the effect of X on M. Likewise, the total effect is the cumulative difference in group-means for Y (Edwards & Lambert, 2007).

This method for testing the mediation hypotheses is favoured over the traditional Baron and Kenny (1986) method for three main reasons. First, the method popularised by Baron and Kenny (1986) suffers from very low statistical power as demonstrated by Fritz and MacKinnon (2007). Second, it has been argued that the traditional Baron and Kenny’s technique overemphasise the importance of a direct effect while suppressing the actual focus of the mediation analysis that is the indirect effect (Zhao, Lynch, & Chen, 2010). Third, the Baron and Kenny (1986) method suffers from the assumption

194

that a lack of correlation between X and Y nullifies the potential for mediation, which has been proven to be a false assumption (Bollen & Stine, 1990; Hayes, Preacher, & Myers, 2011). Zhao et al. (2010) maintain that the test between X and Y is never relevant to establishing mediation. In fact, many researchers (MacKinnon, Krull, & Lockwood, 2000; Preacher & Hayes, 2004; Shrout & Bolger, 2002) indicated that there is no need to demonstrate a relationship to be mediated to establish mediation.

In addition, the Preacher and Hayes (2008) bootstrapping approach is favoured over the classic Sobel test for two reasons. First, the Sobel test is founded on the assumption of normality in the sampling distributions, which has been shown to be a false assumption (Hayes, 2009). Second, the Sobel test is not as statistically robust as compared to a bootstrap test popularised by Preacher and Hayes (2004) and it can only test a single independent variable at a time (Hayes, 2009; Hayes et al., 2011). To test the mediation relationships in this study, Preacher and Hayes (2008) bootstrapping approach was used. Bootstrapping is a widely used technique in social sciences to gauge the extent and significance of indirect effects (Preacher, Rucker, & Hayes, 2007). In bootstrapping, a large number of samples are taken from the data, re-sampling to compute the standard errors of the indirect effect (Preacher & Hayes, 2008). In this study, as recommended by Preacher et al. (2007), 5000 bootstrap samples were used to obtain estimates of the indirect relationships.

The mediation hypotheses H9 to H16 were analysed using the PROCESS macro for SPSS (Model 4, Hayes, 2013). PROCESS calculates a bias-corrected bootstrapped confidence interval (5,000 resamples) for the size of each indirect effect, with significant mediation indicated by a confidence interval that does not contain zero. A macro was downloaded from Hayes’ professional website (http://www.afhayes.com/spss-sas-and-mplus-macros-and-code.html) to conduct the mediation analyses. This macro was added to IBM SPSS Statistics 21.0 to test the

195

proposed mediation hypotheses. The macro allows for the simultaneous testing of several independent variables, mediators and dependent variables and enables the use of the bootstrap method. This study hypothesised that both seeking resources and seeking challenges could mediate the relationship between protean and boundaryless career attitudes and the employee work outcomes. Hence, this study used a parallel multiple mediator models (Hayes, 2013) in which both dimensions of the job crafting behaviour were included as mediators. Specifically, this study tested whether seeking resources and seeking challenges mediated the relationships between four predictor variables (self-directed career management, values-driven career orientation, boundaryless mindset and organisational mobility preference) and four outcome variables (thriving at work, employability, subjective career success and turnover intentions).

Table 5.29 depicts the results of direct and indirect effects analysed using the macro developed by Hayes (2013). The significant indirect paths are indicated by 95% confidence intervals (CI) that exclude zero. In other words, if the lower and upper 95% CIs are either both below or both above zero, there is a statistically significant indirect effect. The results indicate that self-directed career management had positive and significant effects on thriving at work (estimate = .053, SE = .025, 95% CI = [.015, .119]) through seeking challenges, thus supporting Hypothesis 10a. The indirect effect from self-directed career management to thriving at work was non-significant when the mediator was seeking resources (estimate = -.012, SE = .023, 95% CI = [- .063, .032]).

Furthermore, there was positive and significant indirect paths from self-directed career management to employability through seeking resources (estimate = .072, SE = .033, 95% CI = [.021, .153]) and seeking challenges (estimate = .125, SE = .051, 95% CI = [.036, .234]). As such, Hypotheses 11a and 12a were supported. Besides, the indirect effect of organisational mobility preference to employability through seeking resources

196

was positive and significant (estimate = .034, SE = .020, 95% CI = [.005, .092]), thus supporting Hypothesis 11d.

The indirect effect of self-directed career management to turnover intentions through seeking challenges was negative and significant (estimate = -.025, SE = .018; CI = [- .074, -.001]). Thus, Hypothesis 16a was supported. The direct effect between self- directed career management and turnover intentions was not significant (estimate = - .027, SE = .067, t = -.408) indicating full mediation or indirect-only mediation (Zhao et al., 2010).

These findings lend supports to H10a, H11a, H11d, H12a and H16a and suggest that the relationships between self-directed career management and three employee work outcomes (i.e. thriving at work, employability, and turnover intentions) are mediated by seeking challenges. Besides, seeking resources was found to mediate the relationships between self-directed career management and employability, organisational mobility preference and employability. The indirect and direct effect of seeking challenges was significant between self-directed career management and two outcome variables (i.e., thriving at work and employability), suggesting a complementary mediation or partial mediation (Zhao et al., 2010). Similarly, both the indirect and direct effect of seeking resources was significant from self-directed career management and organisational mobility preference to employability, indicating complementary mediations or partial

197

Table 5.29: Summary of Path Models

Path Direct Effects Indirect Effects

Estimate S.E. t-value Estimate S.E. CIL CIU

H9a: SDCM → SR → T 0.551 0.078 7.100*** -0.012 0.023 -0.063 0.032 H9b: VDCO → SR → T -0.291 0.096 -3.034** 0.004 0.012 -0.008 0.047 H9c: BM → SR → T 0.087 0.055 1.572 -0.002 0.006 -0.024 0.005 H9d: OMP → SR → T -0.119 0.061 -1.955 -0.006 0.012 -0.036 0.013 H10a: SDCM → SC → T 0.551 0.078 7.100*** 0.053 0.025 0.015 0.119 H10b: VDCO → SC → T -0.291 0.096 -3.034** -0.0151 0.027 -0.073 0.037 H10c: BM → SC → T 0.087 0.055 1.572 0.0212 0.016 -0.007 0.058 H10d: OMP → SC → T -0.119 0.061 -1.955 -0.024 0.019 -0.006 0.070 H11a: SDCM → SR → PE 0.360 0.087 4.137*** 0.072 0.033 0.021 0.153 H11b: VDCO → SR → PE 0.025 0.107 0.237 -0.025 0.025 -0.078 0.020 H11c: BM → SR → PE -0.013 0.062 -0.205 0.014 0.016 -0.013 0.054 H11d: OMP → SR → PE 0.205 0.068 3.011** 0.034 0.020 0.005 0.092 H12a: SDCM → SC → PE 0.360 0.087 4.137*** 0.125 0.051 0.036 0.234 H12b: VDCO → SC → PE 0.025 0.107 0.237 -0.036 0.062 -0.157 0.088 H12c: BM → SC → PE -0.013 0.062 -0.205 0.050 0.035 -0.023 0.118 H12d: OMP → SC → PE 0.205 0.068 3.011** 0.056 0.041 -0.021 0.144 H13a: SDCM → SR → SCS 0.111 0.049 2.275* 0.026 0.018 -0.002 0.712 H13b: VDCO → SR → SCS 0.064 0.060 1.060 -0.009 0.011 -0.044 0.004 H13c: BM → SR → SCS 0.081 0.035 2.338* 0.005 0.007 -0.003 0.026 H13d: OMP → SR → SCS -0.137 0.038 -3.578*** 0.012 0.010 -0.0002 0.041 H14a: SDCM → SC → SCS 0.111 0.049 2.275* 0.008 0.011 -0.009 0.037 H14b: VDCO → SC → SCS 0.064 0.060 1.060 -0.002 0.007 -0.026 0.005 H14c: BM → SC → SCS 0.081 0.035 2.338* 0.003 0.005 -0.003 0.020 H14d: OMP → SC → SCS -0.137 0.038 -3.578*** 0.004 0.006 -0.003 0.025 H15a: SDCM → SR → TI -0.027 0.067 -0.408 -0.008 0.022 -0.057 0.032 H15b: VDCO → SR → TI 0.273 0.082 3.315** 0.003 0.010 -0.010 0.036 H15c: BM → SR → TI -0.050 0.047 -1.064 -0.002 0.006 -0.021 0.006 H15d: OMP → SR → TI 0.181 0.052 3.456*** -0.004 0.011 -0.035 0.013 H16a: SDCM → SC → TI -0.027 0.067 -0.408 -0.025 0.018 -0.074 -0.001 H16b: VDCO → SC → TI 0.273 0.082 3.315** 0.007 0.015 -0.014 0.047 H16c: BM → SC → TI -0.050 0.047 -1.064 -0.010 0.009 -0.037 0.002 H16d: OMP → SC → TI 0.181 0.052 3.456*** -0.011 0.011 -0.044 0.002

Note. SE = standard error; CIL = lower confidence interval; CIU = upper confidence interval; 5,000

bootstrap samples, *p < .05. **p < .01. ***p < .001. Boldface values represent significant indirect effects. All models include other independent variables as a covariate.

198